Multiscale Topological Properties Of Functional Brain Networks During Motor Imagery After Stroke Fabrizio De Vico Fallania,b,d, Floriana Pichiorria, Giovanni Moronea, Marco Molinaric Fabio Babilonib, Febo Cincottia and Donatella Mattiaa a
Neuroelectrical Imaging and BCI laboratory, IRCCS Fondazione Santa Lucia, Rome, Italy;
b
Department of Physiology and Pharmacology, University Sapienza, Rome, Italy;
c
Experimental Neurorehabilitation laboratory, IRCCS Fondazione Santa Lucia, Rome, Italy;
d
Brain and Spine Institute (CRICM), UPMC/Inserm UMR_S975/CNRS UMR7225, Paris, France
Corresponding author Fabrizio De Vico Fallani, PhD, BME. Email:
[email protected]
Abstract In recent years, network analyses have been used to evaluate brain reorganization following stroke. However, many studies have often focused on single topological scales, leading to an incomplete model of how focal brain lesions affect multiple network properties simultaneously and how changes on smaller scales influence those on larger scales. In an EEG-based experiment on the performance of hand motor imagery (MI) in 20 patients with unilateral stroke, we observed that the anatomic lesion affects the functional brain network on multiple levels. In the beta (13-30 Hz) frequency band, the MI of the affected hand (Ahand) elicited a significantly lower smallworldness and local efficiency (Eloc) versus the unaffected hand (Uhand). Notably, the abnormal reduction in Eloc significantly depended on the increase in interhemispheric connectivity, which was in turn determined primarily by the rise in regional connectivity in the parieto-occipital sites of the affected hemisphere. Further, in contrast to the Uhand MI, in which significantly high connectivity was observed for the contralateral sensorimotor regions of the unaffected hemisphere, the regions that increased in connection during the Ahand MI lay in the frontal and parietal regions of the contralaterally affected hemisphere. Finally, the overall sensorimotor function of our patients, as measured by Fugl-Meyer Assessment (FMA) index, was significantly predicted by the connectivity of their affected hemisphere. These results increase our understanding of stroke-induced alterations in functional brain networks.
Keywords Functional Connectivity, Network theory, EEG, Motor Imagery, Stroke
Abbreviations MRI = Magnetic Resonance Imaging; PET = Positron Emission Tomography; MEG = MagnetoEncephaloGraphy; EEG = ElectroEncephaloGraphy; DTI = Diffusion Tensor Imaging; Ahand = Affected hand; Uhand = Unaffected hand; Ahemi = Affected/Ipsilesional hemisphere; Uhemi = Unaffected/Contralesional hemisphere; MI = Motor imagery; FC = Functional connectivity
1
1 Introduction Most brain functions result from the organization of several neuronal assemblies in a complex and dynamic system (Varela et al., 2001). The term “organization” can be defined as the coherent interdependence of various parts that constitute the whole. Functional connectivity (FC) approaches have been introduced to operationally describe the temporal dependence across spatially remote neurophysiological processors (Friston, 1994); such approaches are effective tools for assessing the organization of the brain, based on the activity of multiple cerebral regions. Over the past decade, graph theory has been introduced as a mathematical approach to characterize the complexity of anatomic and functional brain networks (Bullmore and Sporns, 2009). In functional neuroimaging, a graph is an abstract representation of a pattern of connectivity, in which nodes represent various areas of the brain and links correspond to significant interactions between the activities of regions of the brain. Many groups have exploited graph-based approaches to examine the changes in functional (data-driven) and effective (model-based) connectivity in several brain disorders (He and Evans, 2010). In this regard, many computational studies have focused on understanding how the brain reorganizes its functional structure after stroke from a network-based perspective. Graph theory approaches have allowed the effect of stroke on the organization of the brain to be studied from brain signals that are recorded during resting states and task-related connectivity through various noninvasive techniques, such as functional MRI (fMRI) (Nomura et al., 2010; Wang et al., 2010), EEG (De Vico Fallani et al., 2009), MEG (Westlake et al., 2012), and DTI (Crofts et al., 2011). Although the extent to which the application of such approaches impact the study of stroke-related disturbances in cortical connectivity is unknown, they have been reviewed comprehensively, based on recent meta-analyses (Grefkes and Fink, 2011; Westlake and Nagarajan, 2011). These reviews have highlighted that stroke lesions can effect i) critical deviation from optimal "small-world" network topologies that support processing of segregated and integrated information (Bassett and Bullmore, 2006), ii) altered interhemispheric connectivity, iii) and abnormal region centrality in the ipsilesional and contralesional hemispheres, possibly due to compensatory mechanisms. Although this evidence suggests that stroke modulates several topological attributes of the functional brain network, ranging from small (eg, single-node connectivity) to large scales (eg, connectivity of the entire system), a unifying framework that simultaneously describes the changes in network properties on different scales (Alstott et al., 2009) and their relationships (Vázquez et al., 2004) has not been established and is rarely and poorly applied in analyses of functional brain networks. In this study, we applied a multilevel graph analysis method of functional brain networks that was built from EEG signals and designed to examine multiple topological scales simultaneously. Specifically, based on the peculiarity of functional brain networks that are to be embedded in a physical space that is coincident with the anatomic substrate (Honey et al., 2007; Doron et al., 2012), we aimed to characterize the FC patterns on several scales: i) the entire brain (large scale), ii) the 2 hemispheres (intermediate scale), and iii) each node in the 2 hemispheres (small scale). This framework was adopted to describe the possible brain connectivity disturbances in stroke. Specifically, the functional brain network was studied under a task-specific condition, represented by mental simulation of hand movements, also called motor imagery (MI). MI can be defined as a dynamic state during which the representation of a specific motor action is rehearsed internally without any overt motor output and is governed by the principles of central and peripheral motor control (Decety, 1996). Thus, the practice of action mental imagery
2
by triggering neural activations of relevant brain motor areas is an alternative approach for examining the motor system, even in the absence of movement execution (Page et al., 2007; Sharma et al., 2006). Based on these considerations, we used the proposed graph approach to study the functional brain networks in stroke patients with unilateral cortico-subcortical damage of the sensorimotor system that caused various degrees of motor impairment in the respective contralateral side (ie, hemiplegia or hemiparesis). The patients performed MI with their affected (Ahand) and unaffected (Uhand) hands, the latter of which was used as the reference condition (Jang et al., 2003; Johansen-Berg et al., 2002) to be contrasted to the Ahand MI representing the target condition under investigation. FC was estimated from scalp EEG signals, which have high temporal resolution and carry frequencyspecific information on motor task-related neural activity (Babiloni et al., 1999; Gerloff et al., 1998). We hypothesized that our experimental design would allow us to: assess the impact of unilateral stroke lesions on multiple brain network properties, estimated during the mental rehearsal of movements, and identify possible dependencies between the network changes on various topological scales; and examine the presence of reliable network-based neuromarkers that correlate with poststroke functional motor status, as measured using motor functional scales.
2 Materials and Methods Between 2011 and 2012, we recruited 20 patients (mean age, 55.5 years; 11 females) who were affected by a firstever unilateral stroke in the subacute phase (time since event, 8.4±2.8 weeks) on admission for poststroke rehabilitation treatment at Fondazione Santa Lucia (Rome). All patients had suffered unilateral supratentorial (cortico/subcortical) stroke (left hemisphere 11) that was confirmed by structural MRI and resulted in various degrees of motor impairment on the side of the body that was contralateral to the stroke lesion (for patient details, see Tab. 1). Exclusion criteria were: the pharmacological treatment with drugs affecting the patient’s vigilance and/or the EEG background activity; Mini-Mental State Examination score < 24 (Tombaugh, 2005) and severe cognitive disorders (such as severe hemispatial neglect and language disorders) as evaluated by a neuropsychologist; the presence of other chronic disabling pathologies; orthopedic injuries that could impair reaching or grasping; spasticity of the shoulder, elbow, or finger flexors and extensors that exceeded 3 on the modified Ashworth Scale; The clinical and functional assessment of all patients comprised the following scales: i) the European Stroke Scale (Hantson et al., 1994); ii) the Medical Research Council scale for muscle strength (Compston, 2010) to assess residual strength in the upper limbs; and iii) the upper limb section of the Fugl-Meyer Assessment (Fugl Meyer et al., 1975) to assess functional motor recovery after stroke. Detailed scale scores relative to the clinical and functional assessment are reported in Tab. 1. All measurements were made by an expert physician less than 1 week before the EEG data acquisition. All patients gave written informed consent for participation in the study, which was approved by ethical committee of the Fondazione Santa Lucia.
2.1 EEG Recordings and Motor Tasks All patients had EEGs recorded within 1 week after hospitalization. Patients were comfortably seated in a dimly lit room, with their upper limbs resting on a cushion, and instructed by a visual cue to perform a kinesthetic type of MI of their hand grasping (Jeannerod, 1994). 3
Table 1 - Demographic, clinical, and functional characteristics of the stroke patients
Hand
Lesion
plegia/week
side
F
12
L
51
F
8
L
#3
45
F
12
#4
53
F
#5
41
#6
Patient
Age
Sex
#1
43
#2
Lesion Type
MRC
FMA
ESS
53
18
56
left fronto-temporo-parietal ischemia
45
-
57
R
right fronto- temporo-parietal ischemia
51
-
63
12
R
right temporal and basal ganglia haemorragia
54
-
65
F
8
L
Left-fronto-parietal-rolandic convexity ischemia
77
-
89
47
F
8
R
right fronto-temporo-parietal ischemia
44
-
66
#7
66
F
12
R
right fronto-temporal-parietal ischemia
50
17
63
#8
41
M
12
R
right fronto-temporo-parietal ischemia
49
-
65
#9
64
M
6
R
right nucleo-capsular ischemia
46
10
56
#10
70
M
5
R
right mca thrombosis with ischemia
78
60
96
#11
54
M
7
R
right nucleo-capsular, temporal lobe ischemia
76
49
90
#12
70
M
5
R
left nucleo-capsular, temporal subcortical
46
8
47
#13
57
M
6
R
72
44
78
#14
62
M
12
L
left fronto-temporo-parietal ischemia
72
54
89
#15
64
M
6
L
left fronto-mesial, insular ischemia
70
37
82
#16
71
F
6
L
72
61
76
#17
75
F
10
L
72
44
75
#18
58
M
10
L
semioval center and corona radiata
60
21
66
#19
34
F
8
L
fronto-temporoinsular cortical-subcortical
43
9
62
#20
44
F
4
L
41
5
47
Mean
55,5
8,4
58,6
31,2
69,4
St.Dev.
11,9
2,8
13,4
20,7
14,4
left fronto-parietal,basal ganglia,amigdala ischemia
right emi pons ischemia -
left emipons ischemia left-cortical-subcortical-frontoinsular,prerolandic ischemia
nucleobasal-insular left -
F=Female, M=Male, R=Right, L=Left. ESS = European Stroke Scale: the scale ranges from 0 (maximally affected person) to 100 (normal). MRC = Medical Research Council scale for muscle strength, upper limbs section: the scale ranges from 0 (no movement) to 5 (complete movement against full resistance) for each segment explored (8 segments per side in the upper limb). FMA = Fugl-Meyer Assessment: scores range from 0 (maximally affected) to 66 (normal); FMA was performed in 14 of the 20 recruited patients.
In order to ensure the correct understanding of the MI task by the patients, several trials of actual execution of the same sustained grasping with the unaffected hand were performed before the recording session (visual cue and timing as the EEG experimental condition). Afterwards, in the EEG experiment, patients were instructed to rehearse “the feeling of movements” acquired during the previous MI task practice. Similar pre-EEG recording session practice was allowed with the affected hand by attempting grasping movements. The recording session comprised 2 runs in which the MI of the hand grasping relative to the unaffected (Uhand) and affected (Ahand) hand was sustained for 4 sec. Each run consisted of 30 trials (8 s each), divided equally between randomly presented baseline and task trials. The visual cue was presented using dedicated software, ie,BCI2000 (Schalk et al., 2004), that was synchronized with the EEG amplifiers. 4
As illustrated in Fig. 1, the visual cue was a small red ball that moved at constant speed along the central vertical line of a screen from bottom to top for 8 s (trial duration). In the task trials (panel A), the lower half of the screen was black and the upper half was green. Patients were instructed to be prepared to begin the hand MI as soon as the red ball entered the green area (4 s) and maintain the task until the ball reached the edge of the screen (4 s). In the baseline trials (panel B), the screen was black, and patients simply relaxed throughout the trial duration (8 s). EEG signals were collected from 61 scalp sites that were assembled on an electrode cap per a montage that was modified as an extension of the international 10-20 system. The electroculogram (EOG) was simultaneously recorded to allow the subsequent rejection of ocular artifacts. EEG data were continuously acquired on a commercial system (Brainproduct GmbH, Munich, Germany) with 200 Hz frequency sampling that was referenced to the linked-ear signal. The data were then band-pass filtered in the 1-45 Hz range and depurated from ocular artifacts using the Independent Component Analysis tool (ICA) and commercial software (Vision Analyzer software; Brainproduct GmbH, Munich, Germany).
Figure 1 - Schematic illustration of a representative experimental session. The visual cue is shown on the left side of panel A (Task trial) and B (Rest trial). The right side of panel A and B shows the raw EEG signal recorded from a patient performing the motor imagery of the unaffected hand (Uhand). Traces were obtained from C1 electrode position and were relative to a Task and a Rest trial. In both cases, the trial duration was 8 s and the period of interest for functional connectivity estimation ranged from the 4th to the 8th second.
To ensure that the MI task was performed without any concomitant voluntary muscular contraction, the electromyographic (EMG) activity that was recorded in the left and right opponens pollicis was monitored throughout the experimental session using disposable surface electrodes that were placed in a bipolar belly-tendon configuration. The EMG signals were available to the experimenter to encourage the patients to relax their muscles and avoid movements during the task trials. The preprocessed EEG signals were then segmented considering the last 4 s of each task and baseline trial as the period of interest, as shown in Fig. 1. The segmented traces were inspected visually to reject any EEG segment that had residual muscular or other pronounced artifacts. In the offline analysis, we flipped the functional (EEG time series) and anatomic (scalp electrode positions) data of patients with left-sided lesions along the mid-sagittal plane to perform a group analysis with all 20 patients.
2.2 Functional Connectivity Estimation Brain FC was calculated for the segmented periods of interest (task and baseline) using imaginary coherence (Nolte et al., 2004), which is a robust estimate of synchronization between 2 time series in the frequency domain. This method yields an FC that is unaffected by the volume condition noise due to the anisotropic conductivity of the skull, which blurs the original signals that are generated by the cortical surface. Imaginary coherence gives weighted values 5
between 0 and 1 for each frequency—ie, higher imaginary coherence values in a frequency reflect greater synchronization between the EEG oscillations at that frequency. The original values of imaginary coherence were then Z-transformed to ensure that they approximated a normal distribution (Nolte et al., 2004). To study the level of synchronization in specific physiological frequency contents, the Z-transformed imaginary coherence values were averaged within specific ranges, yielding a single mean value that characterized various EEG bands of interest—namely Theta (4-8 Hz), Alpha (8-13 Hz), Beta (13-30 Hz), and Gamma (30-40 Hz). For each patient, run, and frequency band, the connectivity patterns of the MI task segments were contrasted statistically with those of the baseline segments. The statistical comparison was performed over all possible electrode pairs by paired nonparametric Wilcoxon signed rank test, denoted here as W-test. The functional brain networks that characterized the MI were obtained by maintaining the coherences whose values in the task differed significantly from those at baseline. Similar procedures have been proposed and used in previous studies, in which the functional brain network was the result of a statistical comparison between conditions (Ginestet and Simmons, 2011; Zalesky et al., 2010) or populations (De Vico Fallani et al., 2010). For an analysis of resting states, in which no contrasting procedure is available, the general procedure consisted of repeating the network analysis for a series of increasing threshold values (Rubinov and Sporns, 2010). The significance threshold was set to p=0.05 and adjusted for multiple comparisons by rough false discovery rate (RFDR) correction, based on the number of node pairs for which the W-tests were computed (p adjusted to 0.025). Recently, the RFDR criterion has been applied in several neuroimaging studies (Supekar et al., 2008; Wolf et al., 2011).
2.3 Network Analysis The estimated FC patterns were characterized using network metrics that have been derived from graph theory (Costa et al., 2011). In our graphs, nodes represent scalp electrodes (N=61) and links represent coherence values between pairs of EEG signals. To eliminate any topological bias due to disparate connection densities between brain networks (van Wijk et al., 2010), we transformed the significant coherence values (from the comparison with baseline segments) into binary values and decreased the links in each brain network to the minimum number that was common to all patients, conditions, and frequency bands. Thus, the original sparse weighted graphs were converted into unweighted graphs by retaining only the 185 most significant links (L=185) and transforming them into binary values—ie, 0 = no link and 1 = presence of significant link. Although neglecting the weight of the links (ie, the coherence value) can be considered a reduction of the available information, we noticed that the interpretation and use of link weights in brain network analyses remain an controversial issue, particularly due to the undefined relationship with the concept of physical distance in graphs (Rubinov and Sporns, 2010). A multilevel topological analysis was eventually applied to the estimated brain networks. 2.3.1 Large Scale Two indices were considered to study the more coarsely grained features of the brain networks - global and local efficiency (Eglo and Eloc, respectively) - which have been used extensively to characterize the global properties of functional brain networks (Bullmore and Sporns, 2012). Eglo and Eloc reflect the same properties of the inverse of the average shortest path L and clustering index C, which were introduced by Watts and Strogatz to characterize the small-world property of networks (see Appendix A for details). A simple measure of efficiency-based smallworldness,
6
SW, can be calculated as SW Eloc E
E glo loc r
E glo r
where Elocr and Eglor are the mean efficiency values from equivalent
random graphs (Downes et al., 2012). Whenever SW>1, a network is considered to exhibit small-world properties. 2.3.2 Intermediate Scale The intermediate-scale level was addressed by examining the properties of 2 predetermined sets of nodes, SAhemi and SUhemi, which correspond to the scalp sensors of Ahemi and Uhemi, respectively (see Appendix A for details). Interdensity Kinter is defined as the actual number of links that run between the 2 sets over all possible edges between them. By definition, interdensity ranges from 0 to 1, wherein higher Kinter values reflect a greater number of interhemispheric links. We considered interdensity to be a particular instance of cut size, a graph index that is used frequently to determine the optimal partition, consisting of separate clusters (see Supplementary Text, Section S1.1). Intradensity Kintra is defined as the ratio between the actual number of links in a set and the total number of possible links in the same set. By definition, intradensity ranges from 0 to 1; higher Kintra(S) values indicate greater connection between the nodes in set S. 2.3.3 Small Scale On the small-scale level, we considered graph indices that extracted the finer-grained properties of the EEG network. Two indices were defined to measure the centrality of nodes with respect to the connectivity between and within SAhemi and SUhemi (see Appendix A for details). These sets were symmetric and thus had the same number of nodes NSAhemi=NSUhemi=NS=26 electrodes. Interdegree Dinter was computed as the total number of links of a node in a set to those of the other set. By definition, Dinter ranges from 0 to NS. A node with high Dinter is considered central, because its removal would reduce overall interhemispheric connectivity. Intradegree Dintra was computed as the total number of connections of a node to other vertices in the same group. By definition, it ranges from 0 to NS-1. A node with high Dintra is considered central, because its removal would decrease overall intrahemispheric connectivity. 2.3.4 Normalization by random graphs To handle all normalized network indices, we referred to completely random connectivity patterns, in which links were arranged randomly. Notably, 1000 random graphs were generated by maintaining the same number of nodes and links of the original brain networks. In each instance, links were shuffled randomly without preserving the distribution of node degrees (Sporns and Zwi, 2004). Ultimately, all graph measures that were computed from various brain networks were divided by the respective mean values from the random graphs. When this ratio is lower than 1, the generic brain network property is lower than random graphs; when the ratio exceeds 1, it is higher than random graphs. 2.3.5 Statistical comparison between conditions We use paired nonparametric Wilcoxon signed rank test (W-test), with a statistical threshold of 0.05, to analyze the differences between brain network indices that were computed for Ahand and Uhand conditions (this latter was considered as our reference condition). For small-scale topologies, in which local measures were computed for each node, we corrected for multiple comparisons. Specifically, a nominal significance of p=0.05 was defined and adjusted for multiple comparisons by RFDR correction, based on the number of nodes for which the W-tests were computed (p
7
adjusted to 0.0255). RFDR is a less restrictive procedure for multiple comparisons with greater power than family-wise error rate (FWER) control at the cost of increasing the likelihood of obtaining type I errors (Zar, 1999).
2.4 Interscale dependence between brain network properties In this study, we also examined the dependence of the brain network properties at larger topological scales on those of smaller scales—ie, small -> intermediate, intermediate -> large. The linear regression coefficient was computed between the respective network values (independent variables = larger-scale values, dependent variables = smallerscale values) from all patients under both conditions (Ahand and Uhand) and for each frequency band. Notably, only network attributes that had already reported significant differences between conditions (section 2.3.5) were considered for regression analysis. Evaluation of the regression coefficients determined whether and how changes in finer-grained network properties (smaller scale) influence or “possibly cause” (Sokal and Rohlf, 1994) changes in coarser-grained properties (larger scale). 2.4.1 Intrinsic relationships between graph indices When disparate graph indices are estimated on the same network, they could have a high degree of correlation as a simple consequence of their intrinsic topological definitions—eg, local efficiency and clustering coefficient (Latora and Marchiori, 2001). In this study, we limited such phenomena by considering graph indices that characterized different scales of topology. Nevertheless, these basic relationships should be determined to interpret the interdependence between changes in graph indices fairly. In general, node degrees (intra/inter) are intuitively related to the connection density (intra/inter)—ie, a set of nodes with a higher degree indicates greater connection density. There is a less defined relationship between efficiency values and connection densities. In particular, global (Eglo) and local efficiency (Eloc), which reflect integration and segregation tendencies between groups of nodes, respectively, are intended to be related to the connectivity between hemispheres—ie, Kinter. To examine this issue, we implemented a simulation model that characterized the dynamics of efficiency-based values (SW, Eglo, Eloc) with regard to increasing interhemispheric connectivity (Kinter). The model generated a sequence of networks that had the same size of the brain networks that were considered here—ie, N = 61 and L=185. Starting from a network configuration in which all 185 links were arranged randomly in 1 hemisphere (NS=26 nodes), the model reassigned an increasing number of links between the hemispheres randomly until it reached a configuration that had only interhemispheric links. The choice of such model characteristics was suggested by recent evidence of the effects of stroke on interhemispheric connectivity (Grefkes and Fink, 2011; Westlake and Nagarajan, 2011). To this end, we hypothesized that the estimated brain networks should lie within a range that is delimited by perfect hemisphere lateralization with respect to sensorimotor control of the contralateral hand (McFarland et al., 2000; Volkmann et al., 1998) and abnormal and complete interhemispheric connectivity. Briefly, an increasing number of links l=1,2,...,185 was shuffled randomly in the simulation model. Because we did not know the optimal proportion of the l links to be rearranged between the hemispheres a priori, we introduced a parameter, pinter, to vary the ratio. Thus, the simulated network configuration exhibited interhemispheric links that were proportional to pinter and intrahemispheric links that were proportional to 1-pinter. When pinter=0, we invoked the inferior limit condition in which the l links were reassigned in only 1 hemisphere. When pinter=1, the l links were rearranged exclusively between hemispheres. To simplify this process, 6 equally spaced pinter values were selected: 0, 0.2, 0.4, 0.6, 0.8, and 1. For each 8
pinter value, 1000 random configurations were generated to obtain proper confidence intervals for the simulation model. More details can be found in Appendix B.
2.5 Correlation with functional/clinical measures Statistical correlations were computed between functional/clinical measures in patients and values of the brain network indices that resulted significant after the Ahand versus Uhand contrast (section 2.3.5). Specifically, for those network indices, we considered a delta index (Δ), calculated as the difference between the values for Ahand and Uhand. The Δ values were then used to determine the correlations with functional scale scores. The nonparametric Spearman correlation coefficient R was used to analyze the statistical dependences between the brain network indices and the scores on the functional scales. A p-level of 0.05 was the threshold for statistical significance. This statistical threshold was initially preferred to an adjusted level for multiple comparisons as we wanted to focus on few planned correlations between the Fugl-meyer assessment score (FMA), which is specific of the motor function of the patients, and the significant large-scale and intermediate-scale network indexes that resulted from Ahand versus Uhand. We are aware that this choice is arbitrary and though it reduces family-wise type II errors, it does not control for familywise type I errors (Zar, 1999).
Figure 2 - Grand average (n=20) of EEG network profiles in the Beta band during the MI of the unaffected Uhand and affected Ahand hand. Top part: Grand average of the FC patterns relative to Uhand (panel A) and Ahand (panel B) condition. Blue and red lines denote the links within the unaffected (Uhemi) and the affected (Ahemi) hemisphere, respectively. Gray lines denote the inter-hemispheric links. The intensity of the color and the thickness of the lines vary as function of the number of patients exhibiting that significant link. Bottom part: graph representation of the FC patterns relative to Uhand (panel A) and Ahand
(panel
representation
B)
condition.
nodes
are
In
this
spatially
repositioned through a force-based algorithm so that all the links are approximately of equal length with as few crossing edges as possible. Only links that were in common to more than 4 patients (20% of the sample) are illustrated here. Blue and red nodes indicate scalp electrodes placed over the undamaged (Uhemi) and damaged (Ahemi) hemisphere, respectively. The midline scalp electrodes (from Fpz to Oz) are illustrated as white nodes. Such graph representation highlights the existing partition in different clusters of the Brain networks.
3 Results Fig. 2 shows the grand average (n=20 patients) of the brain networks in the representative Beta band. Under the Uhand condition (panel A), the overall FC tended to converge on the contralateral (contralesional) hemisphere (Uhemi), primarily over the scalp sensorimotor area (electrode C1). In contrast, we did not observe specular behavior in Ahand (panel B), wherein the FC pattern had a similar distribution in the 2 hemispheres with a relatively high representation of frontoparietal and interhemispheric links (grey lines). This profile was also evident in the bottom 9
section of the same figure, in which FC patterns are represented as graphs. Under the Uhand condition, the brain network (represented as a graph) appeared to segregate into 2 primary clusters of nodes, coincident with the 2 hemispheres. In contrast, a more intermingled structure of connectivity emerged in Ahand.
3.1 Multiscale EEG network properties The analysis for the large-scale topological level was performed by comparing the SW, Eglo, and Eloc indices under the Ahand versus Uhand conditions (see 2.3.5). All estimated brain networks tended to have small-world properties—ie, SW>1 (Fig. 3A). In the Beta band, SW was significantly (p=0.025) lower in Ahand compared with Uhand. Similarly, in the Beta band, Eloc under Ahand was significantly (p=0.006) lower than that under Uhand (Fig. 3B). Eglo did not differ between conditions in any EEG frequency band (supplementary Tab. S1). On the intermediate-scale level, we compared inter- and intrahemispheric FC separately between Ahand and Uhand. Interhemispheric connectivity, as measured by the interdensity Kinter, was significantly higher in Ahand for the Beta (p=0.045) band (Fig. 3C). Consistent with this observation, the tendency of the brain networks to form 2 separate clusters that were coincident with the 2 hemispheres was significantly lower (Beta p R=0.475, p=0.0003; CP1 -> R=0.426, p=0.006; CP3 -> R=0.569, p=0.0001) (Fig. 5, upper outer left region). Finally, in the affected hemisphere, there was a significant positive dependence of Kintra(Ahemi) on Dintra of nodes F2, F4 ,and P4 (F2 -> R=0.455, p=0.001; F4 -> R=0.477, p=0.003; P4 -> R=0.547, p=0.0002) (Fig. 5, upper outer right region). No other significant relationships were seen. Figure 4 - Scalp maps of statistical significant differences relative to smallscale network properties in the Beta band. The inter- degree Dinter
statistics are
represented on panel A) and intra-degree Dintra statistics are shown in panel B). Only the significant (p0) condition, while blue colors represent the reversal (Z
